On this pageTimeline of the TestA New Hostname Is RecordedReading the TargetSelecting the TestsTesting Access Between AccountsKeeping a Minimal ProofBranches Checked and LeftThe PT-0218 Finding RecordSOC Analyst ValidationHunting for Earlier UseFix and RetestThe Customer's View of the TestThe Same Run Under Other LimitsWhen the Request Is RefusedAgentic Penetration Testing With ThreatDefenceFrequently Asked Questions This guide follows one such test on the ThreatDefence platform, from a newly recorded hostname to the retest that closed the finding. It covers a single finding on a single host, and the checks selected and the time taken typically differ from one system to the next. Each stage sets out what the agent did, the evidence it kept and the limit that applied. The guide to AI penetration testing explains the method in general terms. Timeline of the TestThe test ran against acme-corp’s customer portal, portal.acmecorp.com, and took under eleven hours from the first record of the hostname to a closed finding. The agents’ own work, from the new hostname to a confirmed weakness, took 17 minutes. Most of the remaining time was the fix. Time Event Done by Record 02:14 portal.acmecorp.com recorded as a new hostname ASM agent Asset record 02:14 to 02:31 Target read, tests selected, test account created Testing agent Run record 02:31 Read access to another account’s invoices proven Testing agent Finding PT-0218, with its proof 02:52 Proof reviewed, severity confirmed, customer notified SOC analyst Finding history 03:10 90 days of web logs searched, no earlier use found Hunter agent Finding history 13:05 Fix reported: access check added to the invoices route acme-corp’s team Finding history 13:11 Same request refused with 403, finding closed Testing agent Finding history Every row has a record behind it that acme-corp’s team and ThreatDefence’s analysts can open. The stages follow in the same order. A New Hostname Is RecordedThe test started because the ASM agent recorded a hostname that was absent from acme-corp’s inventory until then. At 02:14 it added portal.acmecorp.com as a new entry. How discovery finds a hostname that nobody announced is covered in the guide to external attack surface management. A new or changed host is one of the events that starts a test. The Planner agent, which decides what is tested next, placed the hostname ahead of the rolling schedule, and the run opened with its trigger written on it: started by ASM, new hostname recorded. Priority by exposure: The ASM agent records what each hostname serves, and this one answered on port 443 with a sign-in page, so it was queued for testing in the same minute. A parked hostname would have waited for its turn on the schedule. Standing scope: Nobody at acme-corp requested the test, and nothing had to be agreed again before it started. The scope and the limits were already set in the test policy, so a run could open at 02:14 on a host recorded in the same minute. The first record of the test is therefore the asset record itself. Evidence kept: The asset record for portal.acmecorp.com holds the time it first appeared, and the run carries the trigger that started it. Limit applied: Only confirmed and authorized scope is tested, and the hostname qualified because it sits under acmecorp.com, a domain acme-corp had confirmed in its test policy. The excluded host in that policy, pay.acmecorp.com, would have been recorded and left untested. Reading the TargetThe testing agent read what the host serves before it sent any test. The run record shows the reads that shaped the plan. FAIR 64Search this page…secops › acmeacme-corp24hAgent runportal.acmecorp.comstarted by ASM: new hostname recordedReadGET /Customer portal sign-in page. Self-registration is offered.ReadGET /assets/config.jsonAPI base path and route names are listed.PlanTests selectedSelf-registration, access control between accounts, exposed API documentation.TestPOST /api/v1/registerTest account created without approval.TestGET /api/v1/accounts/1042/invoices200 OK. Another account’s records are returned to the test account.ProofLeast intrusive proofResponse headers and one record ID kept. No customer data stored.StopLimit reachedRead access is proven. Anything further waits for approval.Finding raised: any registered account can read another account’s invoices.PT-0218 critical GET /: The host returned a customer portal sign-in page, and the page offered self-registration. GET /assets/config.json: The portal served its configuration file to an anonymous visitor, and the file listed the API base path and the route names. Neither response proves a weakness. Together they told the agent that anyone could ask for an account, and which routes an account could call. The configuration file was read at 02:19, five minutes into the run. Over the whole run the agent read 12 routes in this way and went on to test 3 of them. Evidence kept: The run record holds both requests, what each returned and the agent’s reasoning. Limit applied: Reading uses ordinary requests of the kind a browser sends, and the policy caps them at 5 requests per second per host. Selecting the TestsThe plan came from what was read, so it was specific to this portal. A planning step turned the observations into tests, and worker agents took one test each. Self-registration: The sign-in page offered registration, so the agent would check whether an account could be created without approval. Access control between accounts: The route names included an invoices route addressed by account number, so the agent would check whether one account could reach records that belong to another. Exposed API documentation: The configuration file gave away the API base path, so the agent would check whether a description of the API was published as well. This guide follows the first two, which led to the finding. Practitioners know the second as broken access control and, where a record is fetched by a number in the route, as an insecure direct object reference. The full range of areas the agent can select from is set out in the guide to web application penetration testing. The plan also has an order. The agent starts with the weaknesses attackers use most often against the technology in front of it, and for a customer portal with account numbers in its routes that meant access between accounts. A test of passwords on the sign-in page was left out of the plan, because the policy holds credential tests for approval. Evidence kept: The plan is stored with the observation behind each selected test. Limit applied: The policy for acme-corp holds credential tests and the use of client-supplied accounts for approval, and none of the selected tests fell in either group. Testing Access Between AccountsThe agent proved the weakness by obtaining an account and then reading with it. Registration on the portal was open to anyone, so the agent registered a test account of its own, which needs no approval. Accounts supplied by acme-corp would have waited for one. Request What it tested Response POST /api/v1/register Whether anyone can obtain an account Test account 7731 created without approval GET /api/v1/accounts/1042/invoices, sent with the session of test account 7731 Whether one account can read invoices that belong to account 1042 200 OK, application/json, 3 records The agent’s own account number was 7731. It asked for the invoices of a different number, 1042, and changed nothing else in the request. The expected answer was 403, a refusal, for an account that does not own 1042. At 02:31 the portal returned another account’s records to the test account instead. The test account had been created at 02:27, so the second request followed four minutes after the first. The results form a short path. Open registration alone is a weak control, and the invoices route alone needs an account to reach, but together they put one customer’s records in front of anyone on the internet who registers. Evidence kept: Both requests were written to the run record as they happened, with the status and headers of each response. Limit applied: The policy excludes denial of service, data change and data removal, and the agent registered one account through the portal’s own sign-up and changed no existing record. Keeping a Minimal ProofThe agent kept the smallest proof that shows the weakness is real, then stopped. From the response it retained the headers and one record ID. No customer data was stored, and the invoices stayed where they were. Read access was proven, and that was the agent’s limit. Anything further, such as requesting other account numbers to measure how far the weakness reaches, waits for approval. A proof this small still settles the matter for a developer. The request can be replayed, and the record ID lets acme-corp’s team confirm on their side that the record belongs to account 1042. Evidence kept: The proof consists of the request, the response status and headers, and one record ID. Limit applied: The policy for acme-corp permits read-only proof, the least intrusive kind, and the run ended at the first request that demonstrated the weakness. Branches Checked and LeftThe run touched more of the portal than the route that produced the finding. In 17 minutes the agent sent 14 requests, and the run record shows where it looked and found nothing to report, and where it chose to go no further. Exposed API documentation: The agent asked for a published description of the API at /swagger, and the check ended without a finding. The admin route: A request to /admin was answered with a sign-in requirement, which is the answer the route should give. Going past it would have meant testing credentials, which the policy holds for approval, so the agent recorded the response and left the route. Open registration: The register route was recorded as a weak control, since it issued an account with no approval step. It appears in the run record and in the reasoning behind the severity of PT-0218. Write requests on the invoices route: Whether the test account could also change or delete the invoices of account 1042 was left untested, because the policy excludes data change and data removal. A reviewer can therefore see that the admin route was reached and refused, and can tell a route that was tested from one that was only read. Evidence kept: Each of the 14 requests sits in the run record with its result, including the checks that ended without a finding. Limit applied: Every request went to portal.acmecorp.com, the host in scope for this run, and 14 requests over 17 minutes is far below the cap. The PT-0218 Finding RecordThe run ended by raising finding PT-0218: any registered account can read another account’s invoices. The record covers that one problem on that one host, with enough for a developer to replay it and fix it. FAIR 64Search this page…secops › acmeacme-corp24hFinding PT-0218PT-0218Any registered account can read another account’s invoicesSeverityCriticalExploitabilityConfirmedAssetportal.acmecorp.comStateClosed after retestGET /api/v1/accounts/1042/invoicessession: test account 7731200 OK application/json, 3 recordsexpected: 403 for an account that does not own 1042History02:31Confirmed by the testing agentRead access to another account proven with a test account.02:52Validated by a SOC analystProof reviewed, severity confirmed, your team notified.03:10Hunt completedHunter agent searched 90 days of web logs for the same request pattern. No earlier use found.13:05Fix reportedAccess check added to the invoices route.13:11Retested and closedThe same request is refused with 403. Field Value in PT-0218 Title Any registered account can read another account’s invoices Asset portal.acmecorp.com Severity Critical Exploitability Confirmed Proof GET /api/v1/accounts/1042/invoices with the session of test account 7731 returned 200 OK, application/json, 3 records. Expected: 403 for an account that does not own 1042. Fix Check on the invoices route that the signed-in account owns the account number requested, and refuse with 403 otherwise. State Raised at 02:31, closed after retest at 13:11 History Each later step with its time: validation, hunt, fix report and retest Severity follows exploitability. The weakness was confirmed, it needed nothing beyond an account, and the portal gives an account to anyone who registers. The finding is also the single tracked action item for the problem, so its history and current state sit in one place. Evidence kept: The finding record holds the proof and gathers every later step in its history. Limit applied: A finding is raised only for a weakness the agent has demonstrated, and the proof travels with it. SOC Analyst ValidationA ThreatDefence SOC analyst validated the finding at 02:52, 21 minutes after the agent raised it. The analyst reviewed the proof, confirmed the severity and notified acme-corp’s team. SOC analysts review significant findings, and a critical finding on a customer portal is one. The review works from the recorded request and response, and the agent’s own description of the weakness is checked against them. The record gives the analyst specific things to check before the finding goes out. The accounts differ: The session in the proof belongs to test account 7731 and the route names account 1042, so the request crosses from one account to another. The response is a real read: The portal answered 200 OK with JSON and 3 records. A 200 that carried an empty list or an error page would fall short of proving read access. The method stayed inside the policy: The asset is in scope, the proof is a read, and what was kept is limited to the headers and one record ID. The severity fits the exploitability: The weakness is confirmed, and all it takes to use it is an account that the portal hands out on request. The fix can be acted on: The fix names the route and the check to add, so a developer can start from the finding alone. The agent wrote the finding, and the analyst is the person who decides it is ready for acme-corp’s team. Evidence kept: The validation is entered in the finding history with its time and outcome. Limit applied: A significant finding reaches the customer’s team only after this review, which the 24x7 SOC carried out in the same hour of the night as the finding. Hunting for Earlier UseThe Hunter agent searched acme-corp’s web logs for earlier use of the weakness and found none. The search finished at 03:10 and covered 90 days of logs. It looked for the same request pattern as the proof: the invoices route called for one account number from a session that belongs to another. The test request at 02:31 was already explained by the run record, so the search was for anything before it. What the search needs: A search of this kind depends on logs that record both the route requested and the account behind the session. What a match would have meant: Had the search returned a match, the SOC would have been looking at earlier use of the weakness, with the matching requests and their times as the starting point. This stage depends on monitoring. Because acme-corp is a monitored customer, its web logs were available to search. Where an environment is tested without being monitored, the finding moves to the fix without this step. Evidence kept: The hunt is entered in the finding history with the logs searched, the period, the pattern and the result. Limit applied: The result holds for the logs and the period searched: no matching request appears in 90 days of web logs, and it says nothing about activity outside them. Fix and RetestThe fix was reported at 13:05, and the testing agent’s retest closed the finding at 13:11. Monitoring can be tightened around an affected system while a finding is open, and here the finding was open for under eleven hours. Fix reported. The team at acme-corp marked PT-0218 as fixed, noting that an access check had been added to the invoices route. Retest started. A reported fix is itself a trigger, one of those listed in the guide to continuous penetration testing, so the agent that raised the finding picked it up again. Proof repeated. The agent sent the same request: GET /api/v1/accounts/1042/invoices, with the session of test account 7731. Finding closed. The portal refused the request with 403, the answer expected from the start, and the state changed to closed after retest. Had the portal returned records again, the finding would have been reopened with the new response as fresh evidence. Evidence kept: The retest request and its 403 response sit in the finding history, beside the original 200. Limit applied: The retest repeats the original proof under the same policy as the first run, so it shows that this one request is now refused, and the portal stays on the rolling schedule for everything else. The Customer's View of the TestFor acme-corp’s team the test was mostly quiet. The team was asked to act once, on a finding that had already been proven, validated and checked against the logs. 02:14 to 02:31: Nothing was sent to the team. The asset record and the run could be opened, and the portal’s own web logs recorded the agent’s 14 requests. 02:52: The notification arrived with PT-0218 and its proof, after the analyst’s review. It was the first message of the night. 03:10 onward: The hunt result joined the history, so anyone who opened the finding later read the weakness, the validation and the absence of earlier use in one record. While the finding was open: PT-0218 stayed a single action item with one owner and one current state. 13:05: The team marked the finding as fixed and noted the change, which was its one entry in the record. 13:11: The finding closed, with the 403 from the retest in its history. The work that was absent is part of the picture as well. The team had no test to schedule, no scope to confirm for the new hostname and no report to wait for at the end of an engagement. The guide to penetration testing as a service describes the standing agreement that makes this possible. The Same Run Under Other LimitsThe limits shaped this run mostly by leaving it alone, since the portal was in scope, open to registration and free of a time window. A change to the target or to the policy would have produced a different record. Changed condition What the run would have done Limit behind it The new hostname was pay.acmecorp.com The host would have been recorded as an asset and left untested. pay.acmecorp.com is the excluded host in acme-corp’s policy. The change was on erp.acmecorp.com and was recorded at midday The test would have been held until 22:00. That host has a test window of 22:00 to 05:00. Registration on the portal was closed The reads would have run, and the test between accounts would have waited for accounts from acme-corp and an approval. The use of client-supplied accounts needs approval. A leaked password for a portal account was on record A sign-in test with it would have been queued for an analyst or acme-corp’s team to approve. Credential tests need approval. A page on the portal carried text telling the agent to test another host The request would have been blocked unless that host was in scope. The sandbox’s network policy lets requests reach in-scope targets and nothing else. The approval limit did apply to acme-corp later the same day. At 17:22 a leaked account was matched to acme-corp’s HR portal, and the resulting sign-in test waited for approval. When the Request Is RefusedHad the portal answered 403 at 02:31, the run would have ended with a recorded result and nothing raised. A refusal is the portal working as intended, and a test that the target passes produces a line in the run record. The run record: The reads, the plan, the registration and the refused request would all be logged with their results, as the checks on /swagger and /admin were in the real run. The finding list: Nothing would be added, so there would be nothing for an analyst to validate, for the Hunter agent to search for, or for acme-corp’s team to fix. The asset record: portal.acmecorp.com would stay in the inventory with the time it first appeared, and it would join the rolling schedule. The next test: A later change to the host, a new advisory for a product it runs or a request from the team would start another run. A run without a finding answers a narrower question than it may appear to. It shows that the selected tests were sent that night and what each returned. Tests that were outside the plan, held for approval or excluded by the policy are outside its answer. The record of a quiet run still lets a security lead tell a host that was tested and held up from a host that has yet to be tested. Agentic Penetration Testing With ThreatDefenceThe platform behind this test is ThreatDefence, an AI SOC platform with a 24x7 SOC behind it that automates security monitoring, compliance and cyber risk management. The test followed in this guide is how agentic penetration testing runs on it, and that page lists every limit named in the stages above. The service is available now. The same sequence applies to other internet-facing systems in a confirmed scope: a recorded change, a test that keeps its proof, an analyst’s validation, and a retest once the fix is reported. Where the customer is monitored, the search for earlier use is added. MSPs and MSSPs can deliver the service under their own brand, with a scope set for each client. PT-0218 is the kind of weakness a recorded request and response can prove. Business logic review, social engineering, internal network testing and formal attestation sit outside the service, and an independent human-led test can cover those in parallel. Frequently Asked QuestionsHow long does an agentic penetration test take?In this test, 17 minutes passed between the ASM agent recording the hostname at 02:14 and the testing agent confirming the weakness at 02:31. The finding was closed at 13:11, and most of the time in between was the customer’s fix. Other systems typically take more or less time, depending on what they serve and which tests are selected.What does an agentic pentest report contain?The report is the set of finding records. Each one holds the proof request and the evidence from its response, the asset, the severity, the fix, and a history of every later step with its time. Findings arrive as they are confirmed, so PT-0218 reached the customer’s team the same night it was raised.What does the agent keep as proof of a weakness?The agent keeps the least it needs to show the weakness is real. For PT-0218 that was the request, the response status and headers, and one record ID from the three records returned. No customer data was stored. Every request the agent sent and its result are also logged for the run, so the whole test can be reviewed.Why does the agent stop after one successful request?Proving that the weakness exists is the agent’s task, and one request that returns another account’s records does that. Going further, for example by requesting more account numbers, would read more customer data without adding to the proof. A wider test of that kind runs only after an analyst or the customer’s team approves it.Does a person check the finding before the customer sees it?Yes, for significant findings. A ThreatDefence SOC analyst reviewed the proof for PT-0218 at 02:52, confirmed the critical severity and notified acme-corp’s team. The analyst had the request and the response evidence to work from, because the agent keeps both with the finding.How is earlier use of a weakness checked?For monitored customers, the Hunter agent searches the customer’s telemetry for the request pattern that proved the weakness. For PT-0218 it searched 90 days of web logs and found no earlier use. The answer covers the logs and the period searched, so the range is recorded with the result.What happens when a retest fails?The finding is reopened with fresh evidence. The testing agent repeats the original proof when a fix is reported. If the portal had still returned another account’s records at 13:11, the new response would have been added to the history and the finding would have stayed with the customer’s team.What does the record show when a test finds nothing?The run record still lists each request the agent sent with its result. In this run the checks for published API documentation and on the admin route ended without a finding, and both are recorded beside the request that produced PT-0218. Nothing is raised for a test the target passes, and the host stays on the rolling schedule.What does the customer's team have to do during the test?Very little until a finding arrives. The team at acme-corp was notified once, at 02:52, after the proof had been validated. Its part was to fix the invoices route and mark PT-0218 as fixed at 13:05, which started the retest. The scope, the limits and the approvals had been set in the test policy beforehand.Why was PT-0218 rated critical?Because it was confirmed and simple to exploit, and severity is based on exploitability. The weakness was proven with a real request, it needed no access beyond an ordinary account, and the portal let anyone create an account without approval. In practice that may place other accounts’ invoices within reach of anyone who registers.